Asked to dig through one person’s accounts for forgotten subscriptions, Meta’s Muse turned up $5,350 a year in recurring charges and, as of September 25, 2026, had canceled $1,285 a year of them. That money would otherwise have kept flowing, by default, to services nobody used. Muse is Meta’s consumer AI agent, an assistant that carries out multistep tasks on a user’s behalf, and chores like this one, reading terms of service and wrestling with customer support, look like the place where such agents prove their worth first. Handing an agent anything that touches money still calls for a clear order of operations and a few checkpoints where a person signs off. And not every store will let an agent in: Amazon has already shut Muse out.
What people have handed to Muse
Meta launched Muse on September 8, 2026, and within two weeks it was the top free iPhone app in the US. In its first 12 days it logged 1.8 million downloads and 642,000 daily active users, ahead of ChatGPT’s 1.3 million downloads and 231,000 daily active users over a comparable stretch.
The cases early users have shared are mostly the life admin people put off:
| Task | Result |
|---|---|
| Subscription audit | Found $5,350 a year in recurring charges and canceled $1,285 a year |
| A book subscription a spouse forgot more than a year earlier | Read the refund policy, canceled it and recovered a full year of payments |
| An AT&T customer service call | Spent 98 minutes on hold and through transfers while comparing Verizon and T-Mobile plans, saving more than three hours |
| Uploading an auto insurance policy to find the same coverage for less | Bought a new policy and canceled the old one in about five minutes, saving $3,500 a year |
| Another user’s car insurance | Saved $1,156 a year in under 15 minutes |
| Sports schedules for four children | Merged four schedules spread across four messaging apps into one daily notification |
These are individual results, not savings anyone can count on. The reason this kind of work suits an agent is simple: most people skip refund claims because reading the fine print and calling support feels like more trouble than the money is worth, and an agent absorbs that trouble.
Research suggests subscription businesses lean on exactly that inattention. A study in the American Economic Review, published in May 2025, analyzed payment card records for 10 subscription services from August 2017 to December 2021. It estimated that consumer inertia raises subscription revenue by 87% on average, with a range of 14% to 230% across services. The evidence came from card replacements, which force people to decide again whether to renew and lead to more cancellations. An agent that surfaces unused subscriptions forces the same kind of decision, which could cut into revenue built on inertia.
How to hand off a subscription audit
Muse is designed to be easy for people who have never used an AI tool. Instead of juggling chat threads and projects, you keep one ongoing conversation with a companion character you can name. You can open side chats to set goals or add background, then return to the main conversation without losing context. A sensible sequence:
- Start with phone number verification, and connect accounts such as Gmail or your calendar only when a task like the audit needs them. Muse says it keeps what you choose to share and does not fish for private details.
- Use a side chat to state constraints up front, for example that a service your family shares should stay even if it looks redundant.
- Before anything is canceled, have it check each service’s terms and refund policy, since past charges are sometimes recoverable.
- Watch the task status. Muse labels work as Proposed, which needs your approval, In Progress, and Complete, when the outcome is clear. Do not treat a cancellation or payment as done until it shows Complete.
- When a purchase is involved, connect a payment option such as Link from Stripe that issues single-use virtual cards, so the merchant never sees your real card number.
Cost is unlikely to be a barrier. Ordinary use is free with an allowance of 100 million tokens a week; tokens are the small units of text an AI model processes. Heavier users can pay $20 a month for 500 million tokens or $100 a month for 3 billion. An audit like this should fit within the free allowance. Availability outside the US is still limited, though, because regulations differ from country to country.
▲ Approval, progress and completion stages
What a person should check before approving
Raw model capability now looks like the starting point rather than the differentiator. What matters more is how well the assistant knows your situation. To avoid canceling the wrong contract or booking the wrong date, it has to spot duplicate services, respect constraints such as an account a family member relies on, and remember your earlier corrections. So when a task sits at Proposed, check a few things yourself:
- Is this service truly unused, and does no one else in the household depend on it?
- Is a refund available that a plain cancellation would leave on the table?
- Is the work being done by the AI or by a person?
That last question has a history. Reuters and 404 Media reported that Meta tested having human contractors handle some Muse customer service calls for employees, then rolled the experiment back after concerns about privacy and the lack of disclosure. Users deserve to know clearly who is acting for them.
It also helps to split data access into three questions: what the assistant can reach, what you have allowed, and what it actually uses. A healthy pattern is an assistant that asks for more access step by step as it proves useful. If you are uncomfortable handing personal data to Meta, not using Muse is a perfectly reasonable choice.
Why some stores block shopping agents
Amazon blocked Muse from its store on September 20, 2026. Amazon said Muse did not identify itself as an agent, accessed the store without an agreement and raised concerns about customer credentials. Meta replied that credentials sit in isolated storage the main agent cannot see.
The bigger driver appears to be commercial. Amazon’s advertising revenue rose 22%, from $56.2 billion in 2024 to $68.6 billion in 2025. When shoppers search the store themselves, ads shape what they pick. When an agent decides what to buy before anyone loads an Amazon page, the chance to sell that ad space disappears.
Retailers with smaller ad businesses see it the other way. Walmart’s ad revenue was $6.4 billion in its fiscal year ended January 31, 2026, a 10.7-fold gap with Amazon, so every shopper an agent brings in is new business. Walmart, Best Buy, Gap and Shopify were among the partners Meta announced for Muse at Meta Connect 2026 on September 24. In March 2026, Wired reported that Walmart’s checkout inside ChatGPT converted at about a third of the rate of ordinary click-out links but brought in new customers at roughly twice the rate of other channels. Meta’s plan is to keep Muse free for consumers by default and collect a small transaction fee from merchants when a sale goes through.
For users, this means the same agent may be able to do more at some stores than at others. The economics of agent shopping are still unproven, and subscription and telecom companies whose margins depend on inertia may push back hard against agents that make canceling effortless.
▲ Stores that block agents and stores that welcome them
What to do now
The Muse cases suggest that the first place AI agents earn their keep may be tedious household admin. Anything involving money still needs checkpoints.
- Have an agent review the recurring charges scattered across your card statements and email.
- Ask it to check terms and refund policies before canceling anything.
- Tell it up front which services must stay, such as ones your family uses.
- Review each Proposed action yourself, and after it shows Complete, confirm the cancellation actually went through.
- Connect accounts and payment methods only when needed, and use single-use virtual cards for purchases.
- Check whether the stores you use most allow shopping agents.